Score-Based Conditional Diffusion Model (SCDM)

• SCDM is a conditional generative model that learns score functions from noise-perturbed data and uses reverse-time dynamics to sample from complex conditional distributions.
• The model has been applied successfully in diverse domains such as image synthesis, medical segmentation, and time-series imputation, with performance validated by metrics like DSC and PSNR.
• SCDM design leverages task-driven conditioning mechanisms and various reverse processes (SDEs, ODEs, Langevin dynamics) to tailor its forward and reverse diffusion strategies for specific applications.

Searching arXiv for the primary and related SCDM papers to ground the article in current arXiv records. A score-based conditional diffusion model (SCDM) is a conditional generative model that learns a score field for noise-perturbed targets and then uses reverse-time diffusion, probability-flow ODEs, or Langevin dynamics to sample from a conditional distribution such as $p(x\mid y)$, $p(\theta\mid x_{\mathrm{obs}})$, or $p(\mathbf{x}_0^\mu\mid \mathbf{x}_0^o)$. In the 2021 formulation of conditional image generation, score-based diffusion models are presented as a framework for learning conditional probability distributions, with a systematic comparison of estimators, a theoretical justification for a successful conditional-score estimator, a multi-speed diffusion framework, and the open-source library MSDiff (Batzolis et al., 2021). Subsequent literature uses the same idea across time-series imputation, image synthesis, segmentation, denoising, forecasting, inverse problems, simulator-based inference, and scientific prediction, while also reusing the acronym “SCDM” for distinct named variants in different domains (Tashiro et al., 2021).

1. Formal definition and core equations

At its most generic, an SCDM learns a conditional score

$$s_\theta(x,\sigma,y)\approx \nabla_x \log p_\sigma(x\mid y),$$

where $x$ is the target variable, $y$ is the condition, and $p_\sigma$ is the distribution of $x$ after perturbation at noise level $\sigma$ (Zaman et al., 2023). In continuous-time formulations, the forward process is written as

$$d\mathbf{x}_t=f(\mathbf{x}_t,t)\,dt+g(t)\,d\mathbf{w}_t,$$

and conditional generation uses a reverse-time dynamics driven by the conditional score; one representative form is

$p(\theta\mid x_{\mathrm{obs}})$0

(Cong et al., 27 Aug 2025). In discrete DDPM-style conditional models, the reverse chain is parameterized as $p(\theta\mid x_{\mathrm{obs}})$1, with a network that predicts either noise or a denoised target conditioned on $p(\theta\mid x_{\mathrm{obs}})$2 (Ko et al., 2024).

A second common parameterization replaces the score network by a denoiser $p(\theta\mid x_{\mathrm{obs}})$3. In residual PET/MR denoising, for example, the conditional score is represented as

$p(\theta\mid x_{\mathrm{obs}})$4

so reverse-time sampling operates on the residual variable rather than the full image (Yoon et al., 2024). The same denoiser-to-score relation appears in cortical-thickness trajectory prediction, where

$p(\theta\mid x_{\mathrm{obs}})$5

is used inside an EDM-style reverse ODE (Xiao et al., 2024).

Training is usually framed as denoising score matching. A canonical objective is to minimize the squared error between the learned conditional score and the known score of the forward perturbation kernel, or, equivalently, to regress the injected noise. In CSDI, this becomes a conditional objective for missing-value imputation, explicitly targeting $p(\theta\mid x_{\mathrm{obs}})$6 rather than an unconditional data model (Tashiro et al., 2021).

2. Conditioning mechanisms

The defining feature of an SCDM is not the diffusion process alone but the manner in which conditioning information enters the score field. The condition can be a class label, a semantic map, observed entries in a time series, a previous physical state, a measurement realization, a medical image, a residual predictor, or a simulator output (Tashiro et al., 2021).

In time-series imputation, the condition is the observed subset $p(\theta\mid x_{\mathrm{obs}})$7 together with masks and timestamps, and the model learns a conditional reverse process

$p(\theta\mid x_{\mathrm{obs}})$8

thereby exploiting correlations between observed values (Tashiro et al., 2021). In mechanics inverse problems, the score network is written as $p(\theta\mid x_{\mathrm{obs}})$9, where $p(\mathbf{x}_0^\mu\mid \mathbf{x}_0^o)$0 is a high-dimensional material field and $p(\mathbf{x}_0^\mu\mid \mathbf{x}_0^o)$1 is a noisy measurement field; training uses samples from the joint distribution $p(\mathbf{x}_0^\mu\mid \mathbf{x}_0^o)$2, so a single network can be reused for different measurements after training (Dasgupta et al., 2024). In autoregressive fluid prediction, the condition is the previous physical state $p(\mathbf{x}_0^\mu\mid \mathbf{x}_0^o)$3, encoded and concatenated channel-wise with the noisy current state $p(\mathbf{x}_0^\mu\mid \mathbf{x}_0^o)$4 before U-Net processing (Genuist et al., 30 May 2025).

Conditionality is not restricted to fully paired datasets. OTCS replaces the hard pairing of supervised conditional diffusion by an optimal-transport coupling $p(\mathbf{x}_0^\mu\mid \mathbf{x}_0^o)$5, yielding an OT-guided conditional denoising objective for unpaired or partially paired data (Gu et al., 2023). Nor is conditionality restricted to clean supervision. In semantic image synthesis, the “Stochastic Conditional Diffusion Model” perturbs the conditioning semantic map through a discrete label-diffusion process so that clean and noisy semantic maps become similar as the timestep increases (Ko et al., 2024). In noisy-label image generation, SBDC modifies the effective conditional score at inference time by adding a discriminator-derived correction that approximates the log-likelihood-ratio gradient between clean and corrupted label-conditioned distributions (Cong et al., 27 Aug 2025).

3. Forward-process design and reverse-process variants

Although Gaussian forward corruption remains standard, SCDMs differ markedly in the perturbation process they choose to learn and invert. This design choice determines both what score is estimated and how closely the learned reverse dynamics matches the target task.

Several representative forward processes appear in the literature:

Setting	Forward perturbation	Reverse mechanism
Standard conditional diffusion	Gaussian noising of $p(\mathbf{x}_0^\mu\mid \mathbf{x}_0^o)$6 or $p(\mathbf{x}_0^\mu\mid \mathbf{x}_0^o)$7	Reverse SDE/ODE or DDPM chain
Surf-CDM	Deterministic surface deformation $p(\mathbf{x}_0^\mu\mid \mathbf{x}_0^o)$8	Langevin dynamics on parametric masks
Offshore wind forecasting	Mean-reverting SDE on prediction errors	Reverse SDE on error samples
Digital semantic communications	Channel-matched AWGN-like corruption in constellation space	VE-SDE predictor-corrector denoising

Surf-CDM replaces Gaussian mask perturbation by deterministic “vertical shifting and horizontal rotation” of a parametric surface representation, arguing that the surface re-parameterization makes the score well-defined and smooth while cold-diffusion yields faster reverse convergence (Zaman et al., 2023). In offshore wind forecasting under typhoon conditions, the target variable is the forecasting error, and the forward process is a mean-reverting SDE

$p(\mathbf{x}_0^\mu\mid \mathbf{x}_0^o)$9

with $$s_\theta(x,\sigma,y)\approx \nabla_x \log p_\sigma(x\mid y),$$0, so the model learns the conditional error distribution and reconstructs forecasts by adding sampled errors to a deterministic predictor (He et al., 14 Aug 2025). In PET/MR denoising, the forward process acts on residual volumes $$s_\theta(x,\sigma,y)\approx \nabla_x \log p_\sigma(x\mid y),$$1, not on the clean image itself, and reverse sampling uses a conditional probability-flow ODE in residual space (Yoon et al., 2024). In digital semantic communications, SCDM designs a zero-drift VE forward process to match AWGN corruption of constellation symbols, explicitly avoiding the drift-induced pull toward the origin that would misalign the diffusion process with channel noise (Mo et al., 18 Jan 2025).

The reverse mechanism is equally variable. Some systems use reverse SDEs, some use deterministic probability-flow ODEs, and some use Langevin dynamics. Cortical-thickness trajectory prediction adopts an EDM reverse ODE (Xiao et al., 2024); Surf-CDM uses Langevin dynamics over masks (Zaman et al., 2023); channel denoising uses a predictor-corrector VE sampler (Mo et al., 18 Jan 2025); and simulator-based posterior inference uses reverse SDEs over parameter space (Sharrock et al., 2022). This suggests that “SCDM” denotes a family of conditional score-driven samplers rather than a single canonical algorithm.

4. Representative domains and instantiations

The breadth of SCDM applications is best understood by tracking what plays the role of target $$s_\theta(x,\sigma,y)\approx \nabla_x \log p_\sigma(x\mid y),$$2 and condition $$s_\theta(x,\sigma,y)\approx \nabla_x \log p_\sigma(x\mid y),$$3.

Domain	Condition $$s_\theta(x,\sigma,y)\approx \nabla_x \log p_\sigma(x\mid y),$$4	Target $$s_\theta(x,\sigma,y)\approx \nabla_x \log p_\sigma(x\mid y),$$5
Time-series imputation (Tashiro et al., 2021)	Observed values, masks, timestamps	Missing entries
Medical segmentation (Zaman et al., 2023)	Echocardiogram image	Parametric mask surface
PET/MR denoising (Yoon et al., 2024)	Low-dose PET, MRI, coordinates	Residual volume
Offshore wind forecasting (He et al., 14 Aug 2025)	Typhoon embedding and deterministic forecast	Forecast error
Mechanics inverse problems (Dasgupta et al., 2024)	Noisy mechanical measurements	Material-property field
SBI and posterior sampling (Sharrock et al., 2022)	Observation $$s_\theta(x,\sigma,y)\approx \nabla_x \log p_\sigma(x\mid y),$$6	Parameters $$s_\theta(x,\sigma,y)\approx \nabla_x \log p_\sigma(x\mid y),$$7

In time-series imputation, CSDI reports that probabilistic imputation improves by 40–65% over existing probabilistic imputation methods on popular performance metrics, while deterministic imputation reduces error by 5–20% compared to state-of-the-art deterministic methods (Tashiro et al., 2021). In medical image segmentation, Surf-CDM was evaluated on segmentation of the left ventricle from 65 transthoracic echocardiogram videos (2230 echo image frames) and reported DSC $$s_\theta(x,\sigma,y)\approx \nabla_x \log p_\sigma(x\mid y),$$8, IoU $$s_\theta(x,\sigma,y)\approx \nabla_x \log p_\sigma(x\mid y),$$9, and HD $x$0 (Zaman et al., 2023). In volumetric PET/MR denoising, the CSRD model reported MAE $x$1, PSNR $x$2, SSIM $x$3, $x$4 $x$5, and $x$6 $x$7, and reconstructed a full $x$8 volume within about 3 minutes using 12 GB GPU memory (Yoon et al., 2024).

Outside image processing, the same framework appears in Bayesian inference and scientific modeling. Sequential Neural Posterior Score Estimation treats the posterior $x$9 as the target of a conditional diffusion process on parameters and uses conditional score-based diffusion models to sample from simulator-based posteriors (Sharrock et al., 2022). In inverse problems in mechanics, the score network is trained only from joint samples generated by the forward model, allowing black-box solvers, complex measurement noise, and repeated reuse of the trained model for different observations (Dasgupta et al., 2024). In infinite-dimensional Bayesian linear inverse problems, amortized conditional SDMs are analyzed directly in Hilbert spaces, with a conditional score that can blow up for small times and therefore requires special care absent from the unconditional theory (Baldassari et al., 2023).

5. Theory, misconceptions, and contested points

A first misconception is that “SCDM” denotes a unique standardized architecture. The literature instead uses the term in at least three ways: as a generic label for conditional score-based diffusion; as the specific “Stochastic Conditional Diffusion Model” for semantic image synthesis; and as the “Score-Based Channel Denoising Model” for digital semantic communications (Ko et al., 2024). The underlying unifier is conditional score learning, not a single network or sampler.

A second misconception is that conditioning is equivalent to classifier guidance. Several papers treat exact conditional-score estimation as the central difficulty. Feature-guided score diffusion explicitly avoids direct estimation of the conditional score by introducing a projected score that pushes an image feature vector toward the centroid of the target class in a learned embedding space (Kadkhodaie et al., 2024). SBDC likewise starts from a noisy-label conditional score and adds a discriminator-derived correction term at inference time, arguing that correction is most useful in a mid-noise “conditional phase” where class switching is most likely (Cong et al., 27 Aug 2025). These works suggest that practical conditional generation often relies on score correction or guidance rather than exact modeling of $y$0.

A third misconception is that Gaussian noising is essential. Cold-diffusion segmentation uses deterministic perturbations (Zaman et al., 2023), offshore wind forecasting uses a mean-reverting SDE on errors (He et al., 14 Aug 2025), and channel denoising uses a forward corruption tailored to AWGN in constellation space (Mo et al., 18 Jan 2025). A plausible implication is that SCDM design is often task-driven: the forward process is chosen to make the reverse problem physically or statistically aligned with the domain.

A fourth misconception is that conditional models are inherently supervised by human labels. The analysis in “Why Are Conditional Generative Models Better Than Unconditional Ones?” argues that the key benefit of conditional learning is proper partitioning of the data, and its self-conditioned diffusion model uses k-means clusters in self-supervised feature space as pseudo-conditions, achieving FID 3.94 on ImageNet 64x64 without labels and slightly better FID than the corresponding conditional model on CIFAR10 (Bao et al., 2022). This suggests that the value of conditioning lies in simplifying the conditional subproblems, not necessarily in semantic annotation.

6. Outlook and research directions

Recent work points toward a broader interpretation of SCDM as a conditional transport mechanism rather than merely a conditional image generator. OTCS frames conditional score-based diffusion as transport under an estimated optimal-transport coupling, with a Wasserstein bound linking generated conditionals to the transport plan (Gu et al., 2023). Infinite-dimensional conditional SDMs extend the same logic to discretization-invariant Bayesian inference in function spaces (Baldassari et al., 2023). Feature-guided diffusion shows that low-dimensional class-centroid geometry can serve as a conditioning space even when exact conditional scores are hard to obtain (Kadkhodaie et al., 2024). Fluid-flow prediction shows that a simple conditioned U-Net, combined with careful SDE choice and an energy constraint, can remain stable and physically faithful across multiple scenarios without problem-specific architectural redesign (Genuist et al., 30 May 2025).

Across these strands, SCDMs increasingly appear less as a narrow subclass of diffusion models and more as a general recipe: choose a perturbation process suited to the target variable, learn a score or denoiser conditioned on whatever information narrows the target distribution, and sample by reverse dynamics that respects both the conditional geometry and the task’s noise model. The open problems emphasized in the current literature include exact versus approximate conditional-score estimation, efficient handling of unpaired or noisy conditions, long-horizon stability under autoregressive rollout, and scalable conditional diffusion in high-dimensional or infinite-dimensional settings (Batzolis et al., 2021).